NETWORK ANALYSIS. Duygu Tosun-Turgut, Ph.D. Center for Imaging of Neurodegenerative Diseases Department of Radiology and Biomedical Imaging

Transcription

1 NETWORK ANALYSIS Duygu Tosun-Turgut, Ph.D. Center for Imaging of Neurodegenerative Diseases Department of Radiology and Biomedical Imaging

2 What is a network? - Complex web-like structures

3 Internet is network of routers and computers linked by physical or wireless links

4 Social network Nodes are humans and edges are social relationships

5 Protein-protein interaction networks Network of chemicals connected by chemical reactions

6 Scientific collaboration network

7 Business ties in US biotech-industry

8 Genetic interaction network

9 Ecological networks

10 Graph theory - Study of complex networks - Initially focused on regular graphs - Connections are completely regular, e.g. each node is connected only to nearest neighbors - Since 1950s large-scale networks with no apparent design principles were described as random graphs

11 What makes a problem graph-like? - There are two components to a graph - Nodes and edges - In graph-like problems, these components have natural correspondences to problem elements - Entities are nodes and interactions between entities are edges - Most complex systems are graph-like

12 Graph Theory - History Leonhard Euler's paper on Seven Bridges of Königsberg, published in 1736.

13 Graph Theory - History Cycles in polyhedra Thomas P. Kirkman William R. Hamilton Hamiltonian cycles in Platonic graphs

14 Graph Theory - History Trees in electric circuits Gustav Kirchhoff

15 Graph Theory - History Enumeration of chemical isomers n.b. topological distance a.k.a chemical distance Arthur Cayley James J. Sylvester George Polya

16 Graph Theory - History Four color maps Francis Guthrie Auguste DeMorgan

17 Definition: Graph - G is an ordered triple G:=(V, E, f) - V is a set of nodes, points, or vertices. - E is a set, whose elements are known as edges or lines. - f is a function - maps each element of E - to an unordered pair of vertices in V. - N nodes; connect every pair of nodes with probability p - à Approximately K edges randomly K pn( N 1) 2

18 Definitions - Vertex - Basic element - Drawn as a node or a dot. - Vertex set of G is usually denoted by V(G), or V - Edge - A set of two elements - Drawn as a line connecting two vertices, called end vertices, or endpoints. - The edge set of G is usually denoted by E(G), or E.

19 Example V:={1,2,3,4,5,6} E:={{1,2},{1,5},{2,3},{2,5},{3,4},{4,5},{4,6}}

20 Simple graph is a graph without multiple edges or self-loops.

21 Directed graph (digraph) Edges have directions An edge is an ordered pair of nodes loop multiple edge edge node

22 Weighted graph is a graph for which each edge has an associated weight, usually given by a weight function w: E R

23 Structures and structural metrics - Graph structures are used to isolate interesting or important sections of a graph - Interesting because they form a significant domain-specific structure, or because they significantly contribute to graph properties - A subset of the nodes and edges in a graph that possess certain characteristics, or relate to each other in particular ways - Structural metrics provide a measurement of a structural property of a graph - Global metrics refer to a whole graph - Local metrics refer to a single node in a graph

24 Connectivity - A graph is connected if - you can get from any node to any other by following a sequence of edges OR - any two nodes are connected by a path. - A directed graph is strongly connected if there is a directed path from any node to any other node.

25 Component Every disconnected graph can be split up into a number of connected components.

26 Degree Number of edges incident on a node The degree of 5 is 3

27 Degree (directed graphs) In-degree: Number of edges entering Out-degree: Number of edges leaving Degree = in-degree + out-degree outdeg(1)=2 indeg(1)=0 outdeg(2)=2 indeg(2)=2 outdeg(3)=1 indeg(3)=4

28 Degree: Simple facts If G is a graph with m edges, then Σ deg(v) = 2m = 2 E If G is a digraph then Σ in-degree(v)=σ out-degree(v) = E Number of odd degree nodes is even

29 Walks - A walk of length k in a graph is a succession of k (not necessarily different) edges of the form uv,vw,wx,,yz - This walk is denote by uvwx xz, and is referred to as a walk between u and z. - A walk is closed is u=z. 1,2,5,2,3,4 1,2,5,2,3,2,1 walk of length 5 CW of length 6

30 Path is a walk in which all the edges and all the nodes are different. 1,2,3,4,6 path of length 4

31 Cycle is a closed walk in which all the edges are different. 1,2,5,1 2,3,4,5,2 3-cycle 4-cycle

32 Special types of graphs - Empty graph / Edgeless graph - No edge - Null graph - No nodes - Obviously no edge

33 Trees - Connected acyclic graph - Two nodes have exactly one path between them c.f. routing, later Path Star

34 Regular - Connected graph - All nodes have the same degree

35 Special regular graphs: Cycles C 3 C 4 C 5

36 Bipartite graph - V can be partitioned into 2 sets V 1 and V 2 such that (u,v) E implies - either u V 1 and v V 2 - OR v V 1 and u V 2. - Shows up in coding & modulation algorithms

37

38

39 Complete graph - Every pair of vertices are adjacent - Has n(n-1)/2 edges

40 Complete bipartite graph - Every node of one set is connected to every other node on the other set Stars

41 Planar graphs - Can be drawn on a plane such that no two edges intersect - K4 is the largest complete graph that is planar

42 Subgraph Vertex and edge sets are subsets of those of G a supergraph of a graph G is a graph that contains G as a subgraph.

43 Special subgraphs: Cliques A clique is a maximum complete connected subgraph. A B C D E F G H I

44 Spanning subgraph Subgraph H has the same vertex set as G. Possibly not all the edges H spans G.

45 Spanning tree Let G be a connected graph. Then a spanning tree in G is a subgraph of G that includes every node and is also a tree.

46 Isomorphism - Bijection, i.e., a one-to-one mapping: f : V(G) -> V(H) u and v from G are adjacent if and only if f(u) and f(v) are adjacent in H. - If an isomorphism can be constructed between two graphs, then we say those graphs are isomorphic.

47 Isomorphism Problem - Determining whether two graphs are isomorphic - Although these graphs look very different, they are isomorphic; one isomorphism between them is f(a)=1 f(b)=6 f(c)=8 f(d)=3 f(g)=5 f(h)=2 f(i)=4 f(j)=7

48 Matrix representation of graphs - Incidence Matrix - V x E - [vertex, edges] contains the edge's data - Adjacency Matrix - V x V - Boolean values (adjacent or not) - Or Edge Weights ,6 4,5 3,4 2,5 2,3 1,5 1,

49 List representation of graphs - Edge List - pairs (ordered if directed) of vertices - Optionally weight and other data - Adjacency List (node list)

50 Implementation of a Graph. Adjacency-list representation an array of V lists, one for each vertex in V. For each u V, ADJ [ u ] points to all its adjacent vertices.

51 Edge and Node Lists Edge List Node List

52 Edge lists for weighted graphs Edge List

53 Topological Distance - A shortest path is the minimum path connecting two nodes. - The number of edges in the shortest path connecting p and q is the topological distance between these two nodes, d p,q

54 Distance matrix V x V matrix D = ( d ij ) such that d ij is the topological distance between i and j

55

56 Random graphs & nature - Erdős and Renyi (1959) - N nodes - A pair of nodes has probability p of being connected. N = 12 p = 0.0 ; k = 0 - Average degree, k pn - What interesting things can be said for different values of p or k? (that are true as N à ) p = 0.09 ; k = 1 p = 1.0 ; k ½N 2

57 p = ; k = 0.5 p = 0.0 ; k = 0 p = 0.09 ; k = 1 p = 1.0 ; k ½N 2 - Size of the largest connected cluster - Diameter (maximum path length between nodes) of the largest cluster - Average path length between nodes (if a path exists)

58 Path length L i,j := minimal number of edges that must be traversed to form a direct connection between two nodes i and j L = 1 N 2 L rand i j ln N L i. j ln( K N 1), where K is number of edges

59 p = 0.0 ; k = 0 p = ; k = 0.5 p = 0.09 ; k = 1 p = 1.0 ; k ½N 2 Size of largest component Diameter of largest component Average path length between nodes

60 If k < 1: small, isolated clusters small diameters short path lengths At k = 1: a giant component appears diameter peaks path lengths are high For k > 1: almost all nodes connected diameter shrinks path lengths shorten Percentage of nodes in largest component Diameter of largest component (not to scale) k 1.0 phase transition

61 What does this mean? - If connections between people can be modeled as a random graph, then - Because the average person easily knows more than one person (k >> 1), - We live in a small world where within a few links, we are connected to anyone in the world. - Erdős and Renyi showed that average path length between connected nodes is ln N ln k Erdős and Renyi (1959) Fan Chung David Mumford Peter Belhumeur Kentaro Toyama

62 By definition, for a small world network λ L = 1 L rand

63 α-model Watts (1999) - The people you know aren t randomly chosen. - People tend to get to know those who are two links away (Rapoport, 1957). - The real world exhibits a lot of clustering.

64 Clustering There are cliques or clusters where every node is connected to every other node. Let node i have k i edges which connect it to k i other nodes. K i is number of edges existing between k i nodes. C i = 2K i k i (k i 1) C = i C i C=1 if nearest neighbors of i are also nearest neighbors of other. For a random graph, C rand =p. For a small world networks, γ = C 1 C rand

65 α-model - Add edges to nodes, as in random graphs, but makes links more likely when two nodes have a common friend. - Probability of linkage as a function of number of mutual friends (a is 0 in upper left, 1 in diagonal, and in bottom right curves.)

66 For a range of α values - The world is small (average path length is short) - Groups tend to form (high clustering coefficient). α Clustering coefficient (C) and average path length (L) plotted against a

67 β-model Watts and Strogatz (1998) β = 0 β = β = 1 People know their neighbors. Clustered, but not a small world People know their neighbors, and a few distant people. Clustered and small world People know others at random. Not clustered, but small world

68 Path length and clustering C(0) and L(0) are clustering coefficient and path length for regular graph. For small world, C(p)/C(0) < 1 L(p)/L(0) < 1 Watts and Strogatz, Nature, Vol 393:

69 Empirical Examples of Small World Networks L actual L random C actual C random Film actors Power grid C. elegans Watts and Strogatz, Nature, Vol 393:

70 β-model - First five random links reduce the average path length of the network by half, regardless of N! - Both α and β models reproduce short-path results of random graphs, but also allow for clustering. - Small-world phenomena occur at threshold between order and chaos. Clustering coefficient (C) and average path length (L) plotted against b

71 Power law Albert and Barabasi (1999) - What s the degree (number of edges) distribution over a graph, for real-world graphs? Degree distribution of a random graph, N = 10,000, p = (Curve is a Poisson curve, for comparison.) - Random-graph model results in Poisson distribution.

72 Degree distribution - k i - number of edges connected to a node i - degree of node i

73 Typical shape of a power-law distribution Mathematical relationship: If the frequency (with which an event occurs) varies as a power of some attribute of that event (e.g. its size), the frequency is said to follow a power law. P(k) = λ k e λ k!

74 Many real-world networks exhibit a power-law distribution! WWW hyperlinks Co-starring in movies Co-authorship of physicists Co-authorship of neuroscientists

75 The rich get richer! - Power-law distribution of node distribution arises if number of nodes grow - Edges are added in proportion to the number of edges a node already has - Additional variable fitness coefficient allows for some nodes to grow faster than others

76 Recap - Are real networks fundamentally random? - Intuitively, complex systems must display some organizing principles, which must be encoded in their topology - arrangement in which the nodes of the network are connected to each other

77 Why should we think about the brain as a small world network? - Brain is a complex network on multiple spatial and time scales - Connectivity of neurons - Brain supports segregated and distributed information processing - Somatosensory and visual systems segregated - Distributed processing, executive functions - Brain likely evolved to maximize efficiency and minimize the costs of information processing - Small world topology is associated with high global and local efficiency of parallel information processing, sparse connectivity between nodes, and low wiring costs - Adaptive reconfiguration

78 Three concepts in complex networks Small worlds Clustering Degree distribution

79

80

81

82 How to use network analysis / graph theory to study brain? - Test for small world behavior - Model development or evolution of brain networks - Link network topology to network dynamics (structure to function) - Explore network robustness (vulnerability to damaged nodes, model for neurodegeneration) - Determine if network parameters can help diagnose or distinguish patients from controls - Relate network parameters to cognition

83 Network efficiency and IQ van den Heuvel M P et al. J. Neurosci. 2009;29: The existence of a strong association between the level of global communication efficiency of the functional brain network and intellectual performance - 19 healthy subject - IQ measured with WAIS-III - Resting state fmri - Association was correlation between time-series from each voxel pair (9500 voxels/nodes) - Network constructed for each subject - Network measures were correlated with IQ scores - γ, λ and total connections k - Also correlated normalized path length at each node with IQ

84 Functional network: Small world properties observed for a range of thresholds van den Heuvel M P et al. J. Neurosci. 2009;29:

85 Network parameters vs IQ van den Heuvel M P et al. J. Neurosci. 2009;29:

86 Path length at nodes vs IQ van den Heuvel M P et al. J. Neurosci. 2009;29:

87 Study conclusions - Efficiency of intrinsic resting-state functional connectivity patterns is predictive of cognitive performance - Short path length is crucial for efficient information processing in functional brain networks

88 Cortical Thickness Networks in Temporal Lobe Epilepsy Bernhardt B C et al. Cereb. Cortex 2011;21: patients with drug-resistant TLE; 63 LTLE, 59 RTLE - 47 age- and sex-matched healthy controls - T1-weighted imaging at 1.5T - Estimated cortical thickness in 52 ROIs - Thicknesses corrected for age, gender, overall mean thickness - Computed r ij Pearson product moment cross-correlation across subjects in regions i and j - Explored network parameters at a range of connection densities - Different networks for controls, LTLE, RTLE - Ensures that networks in all groups have the same number of edges or wiring cost - Between group differences reflect topological organization differences; not differences in correlations

89 Cortical thickness correlation networks Bernhardt B C et al. Cereb. Cortex 2011;21:

90 Network parameter analysis of the crosssectional cohort. Bernhardt B C et al. Cereb. Cortex 2011;21:

91 Network robustness analysis. Bernhardt B C et al. Cereb. Cortex 2011;21:

92 Clinical correlation: Relationship between network alterations and seizure outcome after surgery Bernhardt B C et al. Cereb. Cortex 2011;21:

93 A practical framework

94

95

96

97

98

99

100 References - Duncan Watts, Six Degrees (2003). - Albert and Barabasi, Statistical mechanics of complex networks. Review of Modern Physics. 74: (2002) - Kolaczyk, Statistical Analysis of Network Data: Methods and Models, Springer Guye et al., Imaging of structural and functional connectivity: towards a unified definiton of human brain organization? Current Opinion in Neurology 2008, 21: Bassett et al., Hierarchical organization of human cortical networks in health and schizophrenia schizophrenia. The Journal of Neuroscience, 2009, 28(37): Van den Heuvel et al., Efficiency of functional brain networks and intellectual performance. The Journal of Neuroscience, 2009, 29(23): Bassett and Bullmore, Small-World Brain Networks. The Neuroscientist, 2006, 12(6): Bullmore and Sporns, Complex brain networks: graph theorectical analysis of structural and functional systems. Nature Reviews: Neuroscience, 2009, 10: Telesford et al., Reproducilbility of graph metrics in fmri networks. Frontiers in Neuroscience, 2010, 4:article 117.

101 Network / graph analysis software - Brain connectivity toolbox - Matlab BGL -